抹香鯨「尾聲元音」在整段發聲時段內保持穩定
Sperm whale "coda vowels" hold steady through whole bouts
一項預先註冊的分析發現,抹香鯨在同一發聲時段內的相鄰尾聲,其元音標籤大多保持不變,顯示元音更像當下狀態,而非每次重新選擇。1,040 組時段內相鄰尾聲只有 75 組元音不同,少於隨機置換預期的 352.23 組(Q = 0.2129,p = 0.0001)。這項結果限制了把尾聲視為獨立元音樣本的分析,但不能判定狀態來自鯨魚、錄音還是標記方式。
In 2025 and 2026 Begus and colleagues reported that sperm whale codas come in two vowel-like qualities, a and i, that pattern like human vowels (Open Mind 2025; Proc. R. Soc. B 2026). The authors published their hand labels. We asked a question about time: do two neighbouring codas of one whale switch vowel as often as that whale's vowel mix predicts, or does the vowel persist like a state?
We pre-registered the test, published the grader before the job ran, and calibrated it on the real timeline. The labels were joined to recording times in the public Dominica Sperm Whale Project tables, giving 1,134 codas from 13 whales, 19 tag deployments and 94 bouts.
What we found
| test | result |
|---|---|
| within-bout vowel switches (shuffled within whale x deployment x coda type) | 75 observed, 352.23 expected (Q = 0.2129, p = 0.0001) |
| switching by silence before the coda | Q 0.158 at 0-10 s, 0.610 at 10-60 s, 0.909 beyond 300 s |
| same whale, same coda type, different tag deployments | vowel mix differs, p = 0.0001 (4 whales, 10 deployments) |
| calibration | size 0.055; power 1.000 for one vowel per bout with 20% of codas flipped |
| registered predictions | 6 of 7 true (Brier 0.114) |
Only 75 of 1,040 within-bout neighbours differ in vowel. The verdict fixed in advance is VOWELS_CLUSTER_STRONGLY: the label behaves like a state of the moment, not a fresh choice for each coda. The test cannot say which state. A vocal register held through a bout, tag geometry and depth would all produce this. It is a strong constraint on the claim, not proof of an artefact. It does mean that analyses counting codas as independent draws of a vowel overstate the evidence.
How to check
git clone https://github.com/munimtechnologies/whale-vowel-test
cd whale-vowel-test
python -m unittest -v test_vowel_test test_vowel_results
python vowel_test.py --data-dir ./data --out-dir ./out
Read RESULTS-VOWEL.md on GitHub and the pre-registration.
Limits and what comes next
These are the authors' hand labels, made per whale, so the persistence could come from the whale, the recording or the labelling. Only focal-whale codas are labelled, so we could not check whether two whales on one tag share a vowel at the same moment. 1,316 of 1,375 labelled codas matched a unique recording time. The gap profile and the deployment test are side readings. The next check would compare the deployment effect with tag placement and depth records.
Disclosure: this work was carried out with AI agents working for MV2; no human expert has checked it yet. Code Apache-2.0, text CC BY 4.0.
MV2, Munim, Inc., mv2@munimtech.com
來源:DEV Community · dev.to